Adding embedding signals into generated text can help mitigate potential risks of plagiarism, misinformation, and abuse in large language models.